A01Perception and prediction Elucidation of the Mathematical Basis and Neural Mechanisms of Multi-layer Representation Learning

 Reinforcement learning of actions and predictive model learning of the body and the environment require multi-scale representations integrating sensory and motor information. This study aims to elucidate the mathematical basis of deep representation learning and the neural mechanisms of multi-layer representation learning. We further aim to clarify the self-organizing mechanisms for flexible selection and combination of learning modules in the brain.

1) Mathematical basis of multi-layer representation learning

In the framework of deep reinforcement learning, we study how representations that link sensory, motor, and reward information can be effectively learned through computer simulation and mathematical analysis.

2) Neural mechanisms of multi-layer representation learning

In reinforcement learning tasks, we will compare the activity patterns of human brain measured by MRI and those of the hidden layers in deep neural networks to clarify the nature of information coding in different parts of the brain. For the brain areas that turn out to take key roles, we will perform optical neural imaging of corresponding areas in mice to elucidate the neuronal substrates of representation learning.

3) Whole-brain modular self-organization

By integrating the findings from the above, we will explore possible self-organizing mechanisms that enable flexible selection and combination of learning modules in the brain, which include neuromodulation by dopamine and serotonin, and synchronous dynamics of spiking neurons.